一种基于 RGB 比值特征统计模型的高亮点检测算法

摘要

The difficulty of reflectivity inversion for hyperspectral images is of the main obstacle to the application of hyperspectral images. The performance of the commonly used Flat Field method heavily rely on the correct selection of highlight field, unfortunately neither the manual nor automatic selections method performs well in the perspective of accuracy and efficiency. On the basis of extensive experimental collections for hyperspectral images, with the standard referential white board served as ideal white region, the three band in spectrum, i.e. Red, Green and Blue band are chosen in order to exact the ratio characteristics among them. The ratio characteristics are statistically analyzed and demonstrated to be nearly Gaussion distribution, a Gausssion density is thus used to model these characteristics. Subsequently an automatic referential white pixel search method is proposed based on the established Gaussian statistical model, the resultant white pixels lay the foundation of reflectivity inversion for hyperspectral images. Experimental results suggest that the proposed method can effectively improve the accuracy of reflectivity inversion.%高光谱图像反射率反演问题，已成为制约高光谱图像走向应用的重要障碍之一。常用的平场域法关键在于高亮点的正确选取，而目前的人工方法和自动方法均存在选点不准确和效率较低的缺陷。在进行大量的高光谱图像采集实验的基础上，以标准白板图像为基准，对理想白色区域的 R、G、B三个谱段的 DN 值进行了统计分析，用高斯分布拟合了 R、G、B 的比值特征，以此模型为依据，给出了一种基于 R、G、B 谱段 DN 值分析的自动高亮点搜索方法。实验表明，本算法可有效提高高光谱图像反射率反演的准确性。